NatWest: JASPER

Hand JASPER a box of loan documents. Get back a reconciled, structured view of every entity inside, auditable enough for a tier-1 UK bank to act on.

The problem

Commercial lending runs on paper: valuations, leases, accounts, reports. Hundreds of pages per deal, read by expensive people under time pressure. The bank needed machines to do the reading without losing the audit trail.

What we built

A three-tier extraction pipeline that reads the way an analyst does: every content block, then every document, then the whole package, merging what it finds into one deduplicated entity registry. No vector database, no fuzzy retrieval. Structured prompting at every tier, so every answer can show its working.

Built to generalise

Nothing about lending is hard-coded. Entity types arrive as configuration, so pointing JASPER at a new document domain is a prompt exercise, not a rebuild. We maintain a bank-free reference implementation to keep that promise honest.

The numbers

100,000–150,000 tokens per second sustained through Vertex AI across seven regional endpoints, with rate-limit errors held under 0.5%. Running on GKE and Pub/Sub inside one of the most locked-down corporate environments in UK banking.